Correct subcellular localization is fundamental to protein function, and mislocalization is a key driver in diverse human diseases. However, the landscape of localization altering mutations (SAMs) remains largely uncharacterized. Here, we present dbSAM, a comprehensive database curating experimentally validated and computationally predicted SAMs. We manually curated over 500 publications to compile an experimental dataset consisting of 1,512 verirified SAMs and 2,789 mutations with no significant effect on protein localization. By integrating 8,934,325 missense mutations across 18,675 human proteins from public resources, we systematically identified 2,137,705 putative SAMs through a multi-dimensional screening pipeline incorporating two localization prediction models (pSAM and ProtGPS), post-translational modification sites, localization signals, as well as condensate-associated functional domains. We further annotated these variants against disease-related databases, yielding 181,955 pathogenic SAMs implicated in 3,638 Mendelian diseases and 622 cancer types. It indicates that 3%–12% of pathogenic missense mutations trigger protein mislocalization. Our findings highlight protein mislocalization as a prevalent and fundamental pathogenic mechanism underlying human diseases.